MRI DENOISING AND ENHANCEMENT BASED ON OPTIMIZED SINGLE-STAGE PRINCIPLE COMPONENT ANALYSIS

440 views

Published on

Magnetic resonance (MR) images are normally corrupted by Rician noise which may mask the fine details in the image and decreases its resolution. This makes the feature extraction complicated which results in unreliable data analysis. Therefore denoising methods have been applied to improve the image quality. In this study, we propose an optimized single stage principal component analysis (OSPCA) to balance between noise reduction and information preservation. Principal component analysis highlights the data similarities and differences in a better way, which helps in separating noise and edges. Denoising is performed efficiently when the noise and edges are separated. The performance metrics Mean Square Error (MSE), Structural similarity index (SSIM), and Edge Preservation Index (EPI) shows the ability of the proposed method in denoising and preserving information.

Published in: Education
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
440
On SlideShare
0
From Embeds
0
Number of Embeds
2
Actions
Shares
0
Downloads
15
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

MRI DENOISING AND ENHANCEMENT BASED ON OPTIMIZED SINGLE-STAGE PRINCIPLE COMPONENT ANALYSIS

  1. 1. International Journal of Advances in Engineering & Technology, Jan. 2013.©IJAET ISSN: 2231-1963 MRI DENOISING AND ENHANCEMENT BASED ON OPTIMIZED SINGLE-STAGE PRINCIPLE COMPONENT ANALYSIS Amutha .S1, Ramesh Babu D.R2, Ravi Shankar3, Harish Kumar.N4 1, 2, 4 Department of Computer Science & Engineering, Dayananda Sagar College of Engineering, Bangalore, India 3 Department of Information Science & Engineering, Dayananda Sagar College of Engineering, Bangalore, IndiaABSTRACTMagnetic resonance (MR) images are normally corrupted by Rician noise which may mask the fine details in theimage and decreases its resolution. This makes the feature extraction complicated which results in unreliabledata analysis. Therefore denoising methods have been applied to improve the image quality. In this study, wepropose an optimized single stage principal component analysis (OSPCA) to balance between noise reductionand information preservation. Principal component analysis highlights the data similarities and differences in abetter way, which helps in separating noise and edges. Denoising is performed efficiently when the noise andedges are separated. The performance metrics Mean Square Error (MSE), Structural similarity index (SSIM),and Edge Preservation Index (EPI) shows the ability of the proposed method in denoising and preservinginformation.KEYWORDS: Magnetic resonance imaging, Principal Component Analysis, Denoising, Informationpreservation. I. INTRODUCTIONThe quantity of data acquired in Dynamic Contrast Enhancement Magnetic Resonance Imaging(DCE-MRI) is huge. The interpretation of such huge data is a complex task for the radiologist. Toobtain high temporal resolution and to cover a three-dimensional volume over time, short acquisitiontimes are used, resulting in low signal-to-noise ratio (SNR). The accurate detection andcharacterization of small lesions may be crucial [1] because of the poor signal-to-noise ratio. The datais corrupted by hardware -induced noise, geometric distortions, and motion artifacts. So the imagedata should be improved to get good SNR without loss of information. A good denoising algorithmwhich attenuates noise and preserve the fine structural details, with minimum artifacts is essential.Conventional low-pass filters with pixel averaging, can improve the SNR, but they decrease thespatial resolution. The traditional approach like anisotropic diffusion [2] reduces image noise byconsidering a scale space. It is used to suppress the Rician noise. Anisotropic diffusion filtercombined with the Wiener filter [3] has been used for MRI denoising which spatially averages pixelsaccording to their correlation structure. This anisotropic diffusion served well for denoising but finestructures are preserved only to a limited extend. The nonlocal means filter (NLM) has been appliedfor MRI denoising with feature preservation [4,5]. It is based on an estimation of the restored pixelvalue by weighted averaging within an image over a large portion of the pixels.. To preserve theedges, weights are computed by comparing the patches instead of single point. Wavelets have becomean efficient tool in various applications for denoising and data analysis. Wavelet domain denoising ofMR images has produced many algorithms [6, 7]. PCA-based scheme was proposed for imagedenoising by using a moving window to calculate the local statistics, from which the local PCA 224 Vol. 5, Issue 2, pp. 224-230
  2. 2. International Journal of Advances in Engineering & Technology, Jan. 2013.©IJAET ISSN: 2231-1963transformation matrix was estimated [8]. A 3D wavelet transform (3DWT)-based bilateral filtering forRician noise removal was proposed [9]. The delineating capability of 3D WT was used to representthe noisy coefficients. Bilateral filtering of the approximation coefficients in a modified neighborhoodimproved the denoising efficiency with edge preservation. Complex denoising of MR images usingwavelets was proposed by Zaroubi and Goelman [10]. Wavelet has been used for MRI denoising incombination with Radon transform, which estimates noise variance in different scales [11]. OptimalMulticomponent Nonlocal Means PCA (OMNL-PCA) [12] filter reduce noise in the images by usinginformation of the spatial domain as well as in the inter component domain. Unbiased nonlocal meansfilter (UNLM) [13] reduces noise on a set of diffusion weighted images It exploits not only the spatialredundancy, but the redundancy in similar gradient directions as well. It takes into account the Nclosest gradient directions to the direction being processed.The organization of the paper is as follows. In Section II of this paper, the proposed method for MRIdenoising is presented in detail. Section III describes the principal component analysis followed bylocal pixel grouping in section IV. Section V describes the edge enhancement. The evaluation resultsby the MSE, SSIM, and EPI are discussed in Section VI followed by the conclusions in Section VII. II. OPTIMIZED SINGLE-STAGE PRINCIPAL COMPONENT ANALYSIS (OSPCA)In the proposed OSPCA approach, the noisy MRI breast images are taken as input. The first step ofOSPCA is local pixel grouping (LPG). In this, pixels are grouped based on characteristics similarity.A pixel and its nearest neighbors are modeled as a vector variable. The training samples of thisvariable are chosen by grouping the pixels with similar local structures corresponding to theunderlying local window. After grouping the pixels, Principal component analysis transform was usedfor denoising. Due to shrinkage in the PCA domain, the edges are not well preserved even after thelocal statistics of the variables are accurately calculated. Hence an edge enhancement algorithm wasapplied in order to preserve and enhance the edges. Fig. 1 shows the block diagram of the proposedalgorithm OSPCA.III. PRINCIPLE COMPONENT ANALYSISPCA is a classical de-correlation technique in statistical signal processing used mainly in patternrecognition and dimensionality reduction [14]. By transforming the original dataset into PCA domain,the noise and trivial information can be removed by preserving the most significant principalcomponents. T an m-component vector variable and denoted by (1)the sample matrix of y, where yij j=1,2,…..,n, are the discrete samples of variable yi , i=1,2,….,m.The ith row of sample matrix Y, denoted by is called the sample vector of yi.The mean value of Yi is calculated as (2)And the sample vector Yi is centralized matrix of Y is (3)Where = - . Accordingly the centralized matrix of Y is (4) 225 Vol. 5, Issue 2, pp. 224-230
  3. 3. International Journal of Advances in Engineering & Technology, Jan. 2013.©IJAET ISSN: 2231-1963Finally the co-variance matrix of the centralized dataset is calculated as (5)The goal of PCA is to find an orthonormal transformation matrix P to de-correlate . i.e. =P𝑌 so thatthe co-variance matrix of the Z is diagonal. Since the covariance matrix Ω is symmetrical, it can bewritten as (6)Where is the mXm ortho normal eigenvector matrix and is the diagonal eigen value matrix with . The terms and are the eigenvectors and eigen values of .In PCA, the energy of a signal will concentrate on a small subset of the PCA transformed dataset,while the energy of noise will evenly spread over the whole dataset i.e. it fully de-correlates theoriginal dataset , separating signal from noise.Assuming that the original image I is white additive which is corrupted by the Rician noise v, i.e. Iv= I+ v, where Iv is the observed noisy image.An image pixel is described by two quantities, the spatial location and its intensity, while the imagelocal structure is represented as a set of neighboring pixels at different intensity levels. The edgestructures convey its, edge preservation semantic information of an image which is highly desired inimage denoising. In this paper we model a pixel and its nearest neighbours as a vector variable andperform noise reduction on the vector instead of the single pixel.To denoise an underlying pixel, KxK window is centered on it and is denoted by y =[ y1 ….ym]T ,m=K2,the vector containing all the components within the window. The observed image is corruptedby noise.IV. LOCAL PIXEL GROUPING (LPG)LPG procedure is applied in order to group the pixels with similar characteristics for the purpose ofdenoising. For grouping purpose L x L training window is taken. and a K x K variable movingwindow is taken which moves over the training window . Among the techniques for local pixelgrouping we selected block matching because it is simple and more suitable. In Iv which is thecorrupted image there are (L-K+1)2 possible training blocks within L x L training window. Figure 1. Proposed Methodology OSPCA 226 Vol. 5, Issue 2, pp. 224-230
  4. 4. International Journal of Advances in Engineering & Technology, Jan. 2013.©IJAET ISSN: 2231-1963V. EDGE ENHANCEMENT ALGORITHMFig.2 shows the block diagram of the edge enhancement algorithm. The edge information has to beenhanced while the noisy pixels are to be attenuated. To accomplish this we adopt two gainparameters namely G1 and G2. The first gain parameter is used to enhance the edge information whilethe other gain parameter is used to attenuate the noise. These gain parameters are applied based onselecting an arbitrary pixel value such that, the pixel values below the arbitrary pixel selected areassumed to be the edge pixels and the pixels above the arbitrary pixel value are noisy pixels. Thearbitrary pixel in this context is selected by taking minimum and maximum pixel value in the high-pass filtered image and taking the average of both. Extract High frequency component from f(x,y) Select an Arbitrary pixel a(x, y) form f(x,y) If f(x,y) < a(x,y) Considered as Edge pixels. Considered as Noisy pixels Multiply by gain G1 Multiply by gain G2 (G2<G1) + Edge Enhanced image Figure 2. Edge Enhancement AlgorithmVI. EXPERIMENTAL RESULTS AND DISCUSSIONIn this section the performance of the proposed approach is evaluated on DCE MRI breast imagesdownloaded from the National Cancer Institute USA(www. Cancer imaging archive.net). The datasetconsists of T1 weighted and T2 weighted images. For our experimental purpose we used only T2weighted images. The Rician noise of 3% was added to the image. The algorithm OSPCA wasimplemented in MATLAB-7 and applied to the noisy image. The efficiency of the proposed denoisingmethod OSPCA was compared quantitatively and qualitatively with the existing techniques, likeOptimal Multicomponent Nonlocal Means PCA (OMNL-PCA) and unbiased nonlocal means filter(UNLM). For quantitative assessment the mean squared error (MSE) [15], Structural Similarity index 227 Vol. 5, Issue 2, pp. 224-230
  5. 5. International Journal of Advances in Engineering & Technology, Jan. 2013.©IJAET ISSN: 2231-1963(SSIM) [16], and Edge Preservation Index (EPI) [17] were calculated. Table.1 and Fig.3 shows theresults. The results reveal that, our approach is performing better in preserving edges and removingnoise. 6.1 Mean Square Error (MSE)MSE quantifies the deviation of estimated values from the true value. 1 m1 n 1  || I (i, j )  K (i, j) || (7) MSE  2 mn i 0 j 0 6.2. Structural Similarity Index Measure (SSIM)The Structural Similarity Index Measure (SSIM) is used to measure the similarity between twoimages. The SSIM index is a measurement of image quality based on an initial uncompressed ordistortion-free image as reference. SSIM is designed to improve on traditional methods like peaksignal-to-noise ratio (PSNR) and mean squared error (MSE), which have proved to be inconsistentwith human eye perception. The greater value of SSIM denotes the better image quality. (2 y  y + C1 )(2 xy + C2 ) SSIM(x, y) = (  x   y  C1 )( x   y  C2 ) 2 2 2 2 (8)µx :- The average of x.µy :- The average of x.σx2 and σy2 :- variance of x and y.σxy :- The covariance of x and y .C1 and C2 :- Two variables to stabilize the division with weak denominator. 6.3. Edge Preservation Index (EPI)The edge preservation index is defined as follows: (8)Where Io (i, j) is an original image pixel intensity value for the pixel location (x, y), Ip (i, j) is theprocessed image pixel intensity value for the pixel location (x, y). The greater value of EPI gives amuch better indication of image quality. Original image OMNLM_PCA UNLMS Filter Proposed Method OSPCA Original image OMNLM_PCA UNLMS Filter Proposed Method OSPCA 228 Vol. 5, Issue 2, pp. 224-230 Original image OMNLM_PCA UNLMS Filter Proposed Method
  6. 6. International Journal of Advances in Engineering & Technology, Jan. 2013.©IJAET ISSN: 2231-1963 Original image OMNLM_PCA UNLMS Filter Proposed Method OSPCA Figure. 1. The denoising results of 3 MRI images by different Techniques Table 1. MSE, SSIM, and EPI results of the denoised images for noise level of 3% IMAGES METRICS OMNLM_PCA UNLMS PROPOSED METHOD-OSPCA MRI MSE 63.61 27.17 2.8453 Breast SSIM 0.6904 0.6985 0.908 Image 1 EPI 0.7801 0.7467 0.9213 MRI MSE 67.38 80.91 2.2556 Breast SSIM 0.2872 0.5971 0.2891 Image 2 EPI 0.6392 0.6978 0.9032 MRI MSE 26.55 7.9326 2.9319 Breast SSIM 0.5778 0.9292 0.5471 Image 3 EPI 0.7215 0.6819 0.9162V. CONCLUSIONSIn this paper, we have presented a method to improve the signal to noise ratio in MRI breastimages.The proposed method Optimized Single-Stage Principal component analysis (OSPCA)presents a spatially adaptive image denoising method. This method is using single stage principalcomponent analysis, which can protect the image from over blurring. While denoising, the localimage structures have to be preserved. So blocks with similar pixels are gathered using local pixelgrouping (LPG). Then the image is transformed in PCA domain. Shrinking technique is applied on thePCA transformed coefficients to preserve the image fine structures while smoothing noise. Thequantitative performance metrics MSE, SSIM and EPI values are showing better performance whencompared with the existing methods OMNLM_PCA and UNLM. The proposed method can improvethe accuracy of breast cancer diagnosis and detection, especially in the case of subtle tumor.ACKNOWLEDGEMENTSThe authors would like to thank Dr. Sairam Geethanath, Imaging Science Department at ImperialCollege London, for providing the dataset.REFERENCES [1]. Daniel Balvay , Nadjia Kachenoura , Sophie Espinoza , Isabelle Thomassin-Naggara , Laure S. Fournier , Olivier Clement ,Charles-Andr Cuenod , “Signal-to-Noise Ratio Improvement in Dynamic Contrast-enhanced CT and MR Imaging with Automated Principal Component Analysis Filtering” Radiology: Volume 258(2) February 2011. [2]. P Perona, J Malik, Scale-space and edge detection using anisotropic diffusion. IEEE Trans Pattern Anal Mach Intell. 12(7), 629–639 (1990). doi:10.1109/34.56205 [3]. M Martin-Fernandez, C Alberola-Lopez, J Ruiz-Alzola, CF Westin, Sequential anisotropic Wiener filtering applied to 3D MRI data. Magn Reson Imag. 25(2), 278–292 (2007). doi:10.1016/j.mri.2006. [4]. P Coupé, P Yger, S Prima, P Hellier, C Kervrann, C Barillot, An optimized blockwise non local means denoising filter for 3D magnetic resonance images. IEEE Trans Med Imag. 27(4), 425–441 (2008) 229 Vol. 5, Issue 2, pp. 224-230
  7. 7. International Journal of Advances in Engineering & Technology, Jan. 2013.©IJAET ISSN: 2231-1963 [5]. JV Manjón, P Coupé, L Martí-Bonmatí, DL Collins, M Robles, Adaptive nonlocal means denoising of MR images with spatially varying noise levels. J Magn Reson Imag. 31(1), 192–203 (2010). doi:10.1002/jmri.22003 [6]. AM Wink, JBTM Roerdink, Denoising functional MR images: a comparison of wavelet denoising and Gaussian smoothing. IEEE Trans Med Imag. 23, 374–387 (2004). doi:10.1109/TMI.2004.824234. [7]. A Pizurica, AM Wink, E Vansteenkiste, W Philips, J Roerdink, A review of wavelet denoising in MRI and ultrasound brain imaging. Curr Med Imag Rev. 2(2), 247–260 (2006). doi:10.2174/157340506776930665. [8]. Lei Zhang , WeishengDong , DavidZhang , GuangmingShi, “Two-stage image denoising by principal component analysis with local pixel grouping” PatternRecognition43(2010)1531–1549. [9]. Yang Wang1*, Xiaoqian Che2 and Siliang Ma, “Nonlinear filtering based on 3D wavelet transform for MRI denoising” EURASIP Journal on Advances in Signal Processing 2012. [10]. S Zaroubi, G Goelman, Complex denoising of MR data via wavelet analysis: application for functional MRI. Magn Reson Imag. 18(1), 59–68 (2000). doi:10.1016/S0730-725X(99)00100-9. [11]. X Yang, B Fei, “ A wavelet multiscale denoising algorithm for magnetic resonance (MR) images”, . Meas Sci Technol. 22, 025803 (2011). doi:10.1088/ 0957-0233/22/2/025803. [12]. Jos´e V. Manjon, Neil A. Thacker, Juan J. Lull, “Multicomponent MR Image Denoising” doi:10.1155/2009/756897, Volume 2009, International Journal of Biomedical Imaging [13]. Yang Wang, Xiaoqian Che and Siliang Ma, “Nonlinear filtering based on 3D wavelet transform for MRI denoising” EURASIP Journal on Advances in Signal Processing 2012. [14]. Sabita Pal, Rina Mahakud Madhusmita Sahoo, “PCA based Image Denoising using LPG”, 2nd National Conference- Computing, Communication and Sensor Network, CCSN, 2011. [15]. David L Donoho “De-noising by soft thresholding”. IEEE Transactions on Information Theory, 41(3):613–627, May 1995. [16]. Zhou Wang Hamid Rahim Sheikh Eero P. Simoncelli, “Image Quality Assessment: From Error Visibility to Structural Similarity”, IEEE Transactions on image processing,vol.13,no.4,April 2004. [17]. Qin Xue, Zhuanghi Yan, Shengqian Wang ,“Mammographic feature Enhancement Based on Second Generation Curvelet Transform ” International Conference on Biomedical Engineering and Informatics, pp. 349-353, 2010.AUTHORSAmutha S. is pursuing Ph.D in Visvesvaraya Technological University, Karnataka India. Atpresent she is working as an Associate Professor at Dayananda Sagar College of Engineering ,Bangalore. Her research interest areas are image processing and pattern recognition.Ramesh Babu D.R is the HOD & Professor of Computer Science and EngineeringDepartment at Dayananda Sagar College of Engineering, Bangalore. His areas of interestsinclude image processing, pattern recognition and fuzzy theory.RaviShankar is the HOD & Professor of Information Science and Engineering Departmentat Dayananda Sagar College of Engineering, Bangalore. His areas of interests include imageprocessing, pattern recognition and fuzzy theory.Harish Kumar .N is working as Lecturer in Computer Science and Engineering Departmentat Dayananda Sagar College of Engineering, Bangalore. His areas of interests include imageprocessing, segmentation, Pattern Classification. 230 Vol. 5, Issue 2, pp. 224-230

×